The Research On Combining Multiple Classifiers Based On Diversity In Personal Credit Scoring

With China’s commercial banks further strengthening risk managementrequirements, personal credit evaluation index system has constantly been optimized toimprove the ablility in risk management in consumer credit business in order to ensurethe higher demand of the commercial banks. The single classifiers have been constantlyoptimized theoretically and it has reached a high level whether the optimization is fromthe collection of the data, the extraction of the feature or the design of the classifers.Especially under the background of development of the big data, data mining techniquesbecome more sophisticated and personal credit scoring has gradually become one ofvery specific practical problems in this field. And it is in urgent need to select theappropriate theory to reach the core of the personal credit scoring problem. Though alarge number of personal credit evaluation model have been constructed, the modelwhich has the absolute advantage is still undiscovered. In particular, how thetheoretically built model can be applied in pratically is still a big problem. With thecontinuous optimization of the accuracy of the single classifiers and the development ofthe multiple classifier system, this research aims at reaching the further improvement ofthe accuracy and the roubtness in credit scoring.This paper is a systematical study on the research of combining multiple clasifierbased on diversity in credit scoring. This study is under the background of the singleclassfiers becoming more and more, the continuous optimization of them and themultiple classifiers system’s development. Based on these two trends, we propose theidea of combing multiple classifiers model based on the diversity measure to improvethe accuracy and the robustness in personal credit scoring to avoid the loss of personalcredit defaults brought to the commercial banks in their risk management. We firstlydisscuss the design of the model which is combined the multiple classifers based on theD-S evidence theory and the diversity measure. The disscusion is from the aspects ofthe fusion model based on D-S evidence theory, the diversity measure and the concreteidea of how to build the model. We provide some basice concepts and theory in thissection which laid a foundation for the further disscusion in the next two chapters. Thenwe conduct the experiments on combining single classifers based on the D-S evidencetheory and the diversity measure. We applied five single classifers in credit scoring andutilized four diversity measures to evaluate the diversity between them to measure thecomplementarity of them. We select the two which enjoy the biggest diversity andcombine them based on the D-S evidence theory but the result shows that theadvantages of the two single models have not come out in the combing model. So wefurther utilized the accuracy and roubtness to evalutate the single model to sovle thisproblem. Then the result shows that the accuracy and roubtness of the combined model have been improved in this way. Next, we build the credit scoring model based on theensemble theory, in order to optimize the single models and to exclude the impact of thesample structure on the study of the relationship between the accuracy of the fusionmodel and diversity between the single classfiers. We further conduct the diversitymeasure on the the ensemble models and combine the ensemble model based on thediversity measure result and the D-S evidence theory. At last, we conduct the empiricalanalysis in the correlation analysis between the accuracy of the fusion ensemble modeland diversity between the single ensemble classfiers, which aimd at providing atheoretical basis for the guidance of the corresponding follow-up optimization of thepersonal credit scoring models.